[wgsl-in] Allow sign() to take int argument

[fornwall]

Fixes #2458.

The main work here is to make it work on metal, which does not have an int overload of sign(). So we introduce a naga_isign polyfill function there.

Currently the metal backend always emits the naga_isign polyfill. What is a good way to only do that if necessary? We need to declare the function before it's used, so we can't do it after writing out the functions. Is there some mechanism to visit all expressions without duplicating a lot of logic?

[ErichDonGubler]

This mostly LGTM (noting that I don't have merge authority), minus nits and one question: would it be possible to omit this new polyfill unless needed? I'm not familiar with the matter, but I know we've thought about lazily emitting supporting polyfills WRT constructing array members of HLSL structures before. EDIT: 😅 I see that you're also asking this question in the OP:

Currently the metal backend always emits the naga_isign polyfill. What is a good way to only do that if necessary? We need to declare the function before it's used, so we can't do it after writing out the functions. Is there some mechanism to visit all expressions without duplicating a lot of logic?

☝🏻I think @jimblandy or @teoxoy would know the answer to this.

[jimblandy]

Rather than generating a definition for a polyfill, consider simply emitting the code inline. See the way the Metal backend handles MathFunction::Dot for an example:

• It calls back::msl::Writer::put_dot_product to emit the expression at the point of use.
• In back::msl::Writer::update_expressions_to_bake, we ensure that the operand is placed in a temporary variable first, so that we don't evaluate the argument expression more than once.

[fornwall]

Rather than generating a definition for a polyfill, consider simply emitting the code inline.

Thanks for the pointers @jimblandy! Changed to emitting inline in 32e96a5.

[jimblandy]

Thank you very much!